In-group versus Out-group Hardship and Inter-Group Prosociality: A Survey Experiment in Darfur, Sudan

The intervention is a brief reflection prompt administered in Arabic before the outcome module. Two factors are independently randomized, yielding four conditions:

1. Active control: name one ordinary daily activity commonly undertaken in the area.
2. Own-group hardship (T1): name one difficult situation experienced by the respondent and their group over the past year.
3. Other-group hardship (T2): name one difficult situation experienced by members of other groups in the area over the past year; the respondent’s own group is excluded.
4. Both (T3): receive the own-group prompt followed by the other-group prompt, each requiring a separate example.

The four relevant social groups in the context of this study are permanent residents/hosts, displaced people, returnees, and nomads. Respondents’ own-group status is established earlier in the survey and used to tailor the prompts.

All conditions use the same structure, including a request for one concrete example and reassurance that there are no right or wrong answers. The active comparison condition therefore isolates the effect of reflection content rather than reflection itself.

Open-ended responses both facilitate engagement and provide a manipulation check through subsequent coding. Respondents who receive the other-group prompt — that is, those in T2 and in T3 — additionally identify the group they had in mind, allowing verification of the intended treatment and measurement of the specific out-group considered.

The intervention is brief, administered once, and involves no material transfer or programme participation.

2026-08-12

2026-10-31

Commitment to inter-group activity (behavioural, binary). Whether the respondent agrees to sign up for joint activities with members of other groups in the area, recorded as a signature captured on the interviewer's tablet. Coded 1 if the respondent signs, 0 otherwise; reasons are recorded when the respondent declines.

Out-group social distance index (continuous). Mean comfort with members of groups other than the respondent's own across five everyday-interaction domains, following Grady et al. (2023). Higher values indicate lower social distance.

Humanitarian fairness (5-point scale). Agreement with a statement denying other groups equal claim to humanitarian assistance, following Finseraas and Kotsadam (2017). Recorded as 1 = strongly disagree to 5 = strongly agree and reverse-scored, so that higher values are more prosocial.

Competitive victimhood (5-point scale). Agreement with a statement that no other group in the region has suffered as the respondent's own group has, following Bilali and Vollhardt (2013). Recorded as 1 = strongly disagree to 5 = strongly agree and used as recorded, so that higher values indicate more competitive victimhood. Note that this outcome runs in the opposite direction to the three prosociality measures: the treatment effects predicted to be positive for prosociality are predicted to be negative here, and vice versa.

The first three measures form a prosociality family. Competitive victimhood constitutes a second family on its own. False-discovery-rate correction is applied within each family separately, never across them, and the two families are never combined into a single index.

Commitment to inter-group activity. Respondents are asked whether they are willing to take part in joint activities with people from other groups in their area, with a concrete illustration (for example, playing on a mixed team at a sports event). Willingness is recorded only when the respondent physically signs on the tablet, so that expressing interest carries a small but real cost. The variable is 1 for a signature and 0 for any other response, including refusal to sign after a verbal yes.

Out-group social distance index. Comfort is rated separately for each of the four population groups on five items: working in the respondent's field, being paid to watch the respondent's animals, trading goods, sharing a meal, and marrying a close relative. Each item runs 1 (very uncomfortable) to 5 (very comfortable). We first average the five items within each target group to obtain four group-specific scores, then average the three scores that refer to groups other than the respondent's own. This yields one index per respondent, on the original 1–5 metric, capturing average comfort toward out-groups and excluding own-group attitudes. Higher values indicate lower social distance. Group-specific indices are retained as secondary outcomes.

Humanitarian fairness and competitive victimhood. Both are single agree/disagree items, recorded in the instrument as 1 = strongly disagree to 5 = strongly agree, and each is aligned to its own construct rather than to a common direction. The humanitarian fairness statement is worded against the construct — agreeing with it denies other groups an equal claim to assistance — so responses are reverse-scored, giving higher values the more prosocial meaning. The competitive victimhood statement runs with its construct — agreeing with it asserts that one's own group has suffered uniquely — so the recorded response is used as is, giving higher values the meaning of more competitive victimhood. The predicted signs therefore differ between the two: the out-group prompt is expected to raise humanitarian fairness and to lower competitive victimhood.

Units and standardisation. Continuous and rating-scale outcomes are standardised using the mean and standard deviation of the active control cell, so that coefficients are expressed in control-group standard deviations. A single denominator is used for every estimand and every subgroup — main effects, simple effects, the interaction, and all heterogeneity analyses — so that the two main effects and the interaction remain directly comparable and subgroup effects are not confounded with differences in subgroup variance. The binary sign-up outcome is not standardised: it is estimated by linear probability model and reported in percentage points. Raw-scale coefficients and cell means are reported alongside the standardised estimates.

Analysis of the families. The outcomes are analysed separately rather than aggregated into a single index, because the theory makes distinct predictions for prosociality and for competitive victimhood, and because one endpoint is behavioural and binary while the others are attitudinal scales. Anderson's (2008) sharpened q-values (adaptive Benjamini–Krieger–Yekutieli) control the false discovery rate within family; a result counts as significant after correction at q < 0.05, and both adjusted and unadjusted p-values are reported.

Relative hardship comparison. How severely the respondent judges each other group to have been affected by challenges over the past 12 months, relative to their own group; asked for each group other than the respondent's own, on a 5-point scale from much more to much less severely.

Accuracy of beliefs about out-group hardship. Whether the other-group prompt improves the correspondence between a respondent's perceptions of other groups' hardship and objective and self-reported exposure of hardship by members of other groups, rather than only shifting attitudes.

Group-specific social-distance indices. The four target-group-specific scores underlying the primary index, used to examine which out-groups drive any effect.

Cell means, simple effects and pairwise contrasts. Differences of each treatment arm from the active comparison condition, and the contrast between the combined arm and the other-group-only arm.

Manipulation-check measures. Share of respondents giving a valid, on-theme example by arm, and which specific out-group respondents receiving the other-group prompt report having in mind.

Heterogeneity in the primary outcomes by respondent group status and language group, by exposure to conflict and climate shocks, by reported inter-group disputes over land and natural resources, and by zero-sum beliefs, is pre-specified and reported alongside the primary results.

Relative hardship comparison. Each item is analysed on its 5-point scale and also averaged across the three non-own groups to give a single respondent-level score. Coding is aligned so that higher values indicate greater acknowledgement of other groups' hardship relative to one's own. This family receives its own false-discovery-rate correction.

Belief accuracy. Following Eyal, Steffel and Epley (2018), the question is whether making another group's hardship salient changes what people believe to be true, not only how they feel. Each respondent's post-treatment perception of how badly a given out-group was affected is benchmarked against two references: objective exposure at the respondent's location, drawn from conflict event data (ACLED/UCDP), drought indices (SPEI), and flood data (GloFAS); and what members of that group actually report about their own shock exposure in the same sample. Accuracy is the (signed and absolute) discrepancy between perception and benchmark. This outcome sits outside the primary false-discovery-rate families and is reported as a mechanism test.

Manipulation checks. Open-ended examples are coded for whether a valid example was given and whether it is on-theme for the assigned condition (own-group hardship, other-group hardship, or ordinary routine). Coding uses large language models with human verification on a random subset, reported as percent agreement. Primary estimates remain intention-to-treat; any analysis conditioning on on-theme response is exploratory and reported separately.

The study is an individually randomised survey experiment with four arms, structured as a 2×2 factorial and embedded in a face-to-face survey covering roughly 3,000 respondents in about 150 villages across North, Central and South Darfur (approximately 20 respondents per village).

Two reflection prompts are crossed as independent on/off factors: a prompt making one's own group's hardship salient, and a prompt making other groups' hardship salient. Crossing them produces four cells — neither prompt (an active comparison reflection task), own-group prompt only (T1), other-group prompt only (T2), and both in sequence (T3) — with equal assignment probability, so roughly 750 respondents per cell. Assignment is at the individual respondent level and occurs inside the survey software.

Because the two prompts are expected to move outcomes in opposite directions, they are hence not pooled into a single treatment indicator. The pre-specified estimands are the two factorial main effects — the effect of switching each prompt on, averaged over whether the other prompt is present, each estimated on the full sample — and their interaction, which tests whether receiving both departs from additivity. Cell means, simple effects against the comparison condition, and the contrast between the combined arm and the other-group-only arm are reported as secondary descriptive quantities.

Estimation is by analysis of covariance. Outcomes are regressed on the two treatment factors and their product, a pre-specified vector of covariates measured before treatment, and village and enumerator fixed effects. The two factors are effect-coded (−½ when the prompt is absent, +½ when present), so that the two reported coefficients are the factorial main effects — each the difference between prompt-present and prompt-absent cells, averaged over the other factor and estimated on all 3,000 respondents — and the product coefficient is the difference-in-differences. Enumerator fixed effects are included because prompts are read aloud and delivery may vary. Both adjusted and unadjusted estimates are reported. Standard errors are heteroskedasticity-robust and clustered at the village level in the main specifications; where the social-distance data are analysed in long form at the item-by-target level, standard errors are two-way clustered at the respondent and village levels. False discovery rate is controlled within outcome family using sharpened q-values.

All covariates and moderators are measured before the reflection module. Balance is assessed by reporting covariate means by arm together with an omnibus test.

Not available

Randomisation is done inside the survey instrument (administered via SurveyCTO on tablets), at the moment of interview. Assignment is generated by hidden calculate fields that are not visible to either the respondent or the enumerator, so neither party knows the arm in advance and the enumerator cannot select it. Each of the two factors — the own-group hardship prompt and the other-group hardship prompt — is switched on independently with probability one-half, which is equivalent to assigning each respondent with equal probability 1:1:1:1 to the four arms.

The individual respondent. There is a single level of randomisation: within each village, individual respondents are independently assigned to the four arms. Villages and households are sampling units, not units of randomisation — every arm appears within every village. There is no group-, household- or village-level assignment.

No

Randomisation is not clustered, so the units of randomisation are the 3,000 individual respondents themselves.

For reference on the sampling structure: respondents are drawn from approximately 150 villages in North, Central and South Darfur, with about 20 households per village. Villages are sampling and standard-error clusters, not randomisation clusters — all four arms are represented within each village.

3,000 individual respondents (one adult respondent per sampled household), across approximately 150 villages in three Darfur states.

Equal allocation across four arms, approximately 750 respondents each:

750 respondents — active comparison condition (neither prompt)
750 respondents — own-group hardship prompt only
750 respondents — other-group hardship prompt only
750 respondents — both prompts, administered in sequence
Equivalently, by factor: 1,500 respondents receive the own-group prompt and 1,500 do not; 1,500 receive the other-group prompt and 1,500 do not.

Calculations assume 3,000 respondents in four equal arms of 750, 80% power and a two-sided test at alpha = 0.05. Rating-scale effects are expressed in standard deviations of the outcome; effects on the binary sign-up outcome are in percentage points, assuming a 50% agreement rate, which is the highest-variance and therefore most conservative case. The design detects main effects of roughly 0.10 SD (about 5 percentage points), or 0.12 SD (about 6 points) under the stricter post-correction threshold, and interactions of roughly 0.21 SD (about 10 points). Precision differs by estimand because the effective sample differs. Each main effect compares all 1,500 respondents who received a given prompt with the 1,500 who did not, and is best powered. A simple effect compares one 750-respondent cell with another. The interaction is a difference between two differences and requires roughly twice the effect size to detect; it is the least well-powered pre-specified quantity. The post-correction figures apply a conservative Bonferroni threshold (two-sided α = 0.05/3, reflecting the three outcomes in the prosociality family) rather than the sharpened q-values used in analysis. Because the adaptive Benjamini–Krieger–Yekutieli procedure is less conservative than Bonferroni, these are an upper bound on the effect sizes detectable after correction. No design-effect deduction is taken for clustering, and none is required. Randomisation is at the individual level within village, so all four arms appear in every village and the arms are balanced within each village; within-village correlation in outcomes therefore costs no precision in the way it would under village-level assignment. Village and enumerator fixed effects and the pre-treatment covariates absorb further outcome variation, so realised precision should be at least as good as these figures imply. No attrition allowance is made, since all outcomes are collected in the same interview as the treatment.